Add streaming data pipeline and giant CLI entry point

- Streaming pipeline: row-group-level parquet reading (PyArrow) so
  large files never fully land in RAM; Welford online algorithm for
  normalizer fitting; StreamingStepsDataset with shuffle buffer and
  multi-worker file striping; event-ID scan and vocab scan via cheap
  single-column reads
- giant/cli.py: typer-based CLI with `giant train` subcommand, mirroring
  scripts/train.py; --shuffle-buffer flag for RAM control
- pyproject.toml: add typer>=0.12 dependency and giant entry point
- train.py: replace len(loader.dataset) with local counters (compatible
  with IterableDataset which has no __len__)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
2026-06-17 11:03:48 +02:00
parent 9277d79dff
commit 93c4d6b74d
8 changed files with 538 additions and 30 deletions
+202
View File
@@ -0,0 +1,202 @@
import subprocess
import tomllib
from enum import Enum
from pathlib import Path
from typing import Optional
import numpy as np
import torch
import typer
from torch.utils.data import DataLoader
from typing_extensions import Annotated
from giant.data.loader import (
find_parquet_files,
load_event_ids,
iter_file_chunks,
build_index_maps_from_files,
)
from giant.data.transforms import build_features, _WelfordAccumulator
from giant.data.dataset import make_event_split, StreamingStepsDataset
from giant.model.network import DenoisingMLP
from giant.train import train as run_training
app = typer.Typer(no_args_is_help=True)
@app.callback()
def _main() -> None:
"""GIANT — Geant4 step-function surrogate."""
class Mode(str, Enum):
flow = "flow"
ddpm = "ddpm"
def _auto_device() -> torch.device:
if torch.cuda.is_available():
return torch.device("cuda")
if torch.backends.mps.is_available():
return torch.device("mps")
return torch.device("cpu")
def _git_hash() -> str:
try:
return subprocess.check_output(
["git", "rev-parse", "HEAD"], stderr=subprocess.DEVNULL
).decode().strip()
except Exception:
return "unknown"
def _load_toml(path: Path) -> dict:
with open(path, "rb") as f:
return tomllib.load(f)
def _save_config(cfg: dict, out_dir: Path) -> None:
lines = [f"# git: {_git_hash()}", ""]
for section, values in cfg.items():
lines.append(f"[{section}]")
for k, v in values.items():
lines.append(f"{k:<12} = {repr(v) if isinstance(v, str) else v}")
lines.append("")
(out_dir / "config.toml").write_text("\n".join(lines))
@app.command()
def train(
data: Annotated[Path, typer.Argument(help="Parquet file or directory of parquet files")],
config: Annotated[Optional[Path], typer.Option(help="TOML config file (overridden by explicit flags)")] = None,
mode: Annotated[Optional[Mode], typer.Option(help="Generative model: flow matching or DDPM")] = None,
epochs: Annotated[Optional[int], typer.Option()] = None,
batch_size: Annotated[Optional[int], typer.Option()] = None,
lr: Annotated[Optional[float], typer.Option()] = None,
hidden_dim: Annotated[Optional[int], typer.Option()] = None,
n_blocks: Annotated[Optional[int], typer.Option()] = None,
emb_dim: Annotated[Optional[int], typer.Option()] = None,
val_fraction: Annotated[Optional[float], typer.Option()] = None,
shuffle_buffer: Annotated[int, typer.Option(help="Rows held in RAM per worker for shuffling")] = 65536,
out: Annotated[Optional[Path], typer.Option(help="Checkpoint dir (default: auto from hyperparams)")] = None,
device: Annotated[Optional[str], typer.Option(help="cpu | cuda | mps (default: auto)")] = None,
num_workers: Annotated[Optional[int], typer.Option()] = None,
) -> None:
"""Train the GIANT surrogate model."""
# Defaults → config file → explicit CLI flags.
cfg: dict = {
"train": {
"mode": "flow", "epochs": 100, "batch_size": 4096, "lr": 3e-4,
"val_fraction": 0.1, "num_workers": 4,
},
"model": {
"hidden_dim": 256, "n_blocks": 6, "emb_dim": 16,
},
}
if config is not None:
file_cfg = _load_toml(config)
for section in ("train", "model"):
cfg[section].update(file_cfg.get(section, {}))
cli_train = {k: v for k, v in {
"mode": mode.value if mode is not None else None,
"epochs": epochs, "batch_size": batch_size, "lr": lr,
"val_fraction": val_fraction, "num_workers": num_workers,
}.items() if v is not None}
cli_model = {k: v for k, v in {
"hidden_dim": hidden_dim, "n_blocks": n_blocks, "emb_dim": emb_dim,
}.items() if v is not None}
cfg["train"].update(cli_train)
cfg["model"].update(cli_model)
t, m = cfg["train"], cfg["model"]
_device = torch.device(device) if device else _auto_device()
out_dir = out or Path(
f"checkpoints/{t['mode']}"
f"_h{m['hidden_dim']}"
f"_b{m['n_blocks']}"
f"_e{m['emb_dim']}"
f"_lr{t['lr']}"
f"_bs{t['batch_size']}"
)
typer.echo(f"device: {_device}")
typer.echo(f"out_dir: {out_dir}")
files = find_parquet_files(data)
typer.echo(f"found {len(files)} parquet file(s)")
typer.echo("scanning event IDs …")
all_event_ids = np.concatenate([load_event_ids(f) for f in files])
train_events, val_events = make_event_split(all_event_ids, val_fraction=t["val_fraction"])
events_arr = np.array(sorted(train_events))
n_train_steps = int(np.isin(all_event_ids, events_arr).sum())
typer.echo(
f" {len(all_event_ids):,} steps | "
f"{len(train_events)} train events (~{n_train_steps:,} steps) | "
f"{len(val_events)} val events"
)
typer.echo("building vocabulary maps …")
pdg_map, mat_map = build_index_maps_from_files(files)
typer.echo(f" {len(pdg_map)} PDG codes | {len(mat_map)} materials")
typer.echo("fitting normalizer (streaming) …")
cond_acc = _WelfordAccumulator(9)
tgt_acc = _WelfordAccumulator(6)
for path in files:
for chunk in iter_file_chunks(path):
mask = np.isin(chunk["event_id"], events_arr)
if not mask.any():
continue
chunk_tr = {k: v[mask] for k, v in chunk.items()}
cond_cont, _, target, _, _ = build_features(chunk_tr, pdg_map, mat_map)
cond_acc.update(cond_cont)
tgt_acc.update(target)
cond_norm = cond_acc.to_normalizer()
tgt_norm = tgt_acc.to_normalizer()
train_ds = StreamingStepsDataset(
files=files, split_events=train_events,
pdg_map=pdg_map, mat_map=mat_map,
cond_normalizer=cond_norm, target_normalizer=tgt_norm,
shuffle_buffer=shuffle_buffer, shuffle=True,
)
val_ds = StreamingStepsDataset(
files=files, split_events=val_events,
pdg_map=pdg_map, mat_map=mat_map,
cond_normalizer=cond_norm, target_normalizer=tgt_norm,
shuffle=False,
)
pin = _device.type == "cuda"
train_loader = DataLoader(
train_ds, batch_size=t["batch_size"],
num_workers=t["num_workers"], pin_memory=pin,
)
val_loader = DataLoader(
val_ds, batch_size=t["batch_size"],
num_workers=t["num_workers"], pin_memory=pin,
)
model = DenoisingMLP(
pdg_vocab=len(pdg_map), mat_vocab=len(mat_map),
hidden_dim=m["hidden_dim"], n_blocks=m["n_blocks"], emb_dim=m["emb_dim"],
)
typer.echo(f"model: {sum(p.numel() for p in model.parameters()):,} parameters")
out_dir.mkdir(parents=True, exist_ok=True)
_save_config(cfg, out_dir)
run_training(
model=model,
train_loader=train_loader, val_loader=val_loader,
mode=t["mode"], epochs=t["epochs"], lr=t["lr"],
device=_device, out_dir=out_dir,
normalizer_dict={"cond": cond_norm.to_dict(), "target": tgt_norm.to_dict()},
pdg_map={str(k): v for k, v in pdg_map.items()},
mat_map={str(k): v for k, v in mat_map.items()},
)